Option one: filters and elbow grease
Most practitioners still run the same playbook they ran ten years ago. Bloomberg, Capital IQ, FactSet. SIC or GICS filter, geography filter, market cap band, sanity-check the list by eye, drop the outliers. What it does have going for it: the initial screening is a decision you can point to. Ask why a company is or isn't on the list, and the answer is a filter setting, not a feeling. Then comes the judgement, reading through company descriptions, looking at financials, choosing your top picks. Slow, manual work, enough that most of the actual selection logic ends up living in an analyst's head rather than on paper.
Option two: the AI blackbox search
Today, a user can drop a business description into ChatGPT or Claude and ask for comparable companies, and it will hand you eight or ten names. No filter, no screen, no visible logic. The model is pattern-matching against whatever it absorbed in training, which means familiar, frequently-discussed companies show up disproportionately, and the same question can return a different list five minutes later. Useful as a gut check. Not something you could reconstruct or defend if someone later asked how the list was built. If you ask different models, or at a different point in time, you are likely to get different results. Now, yes, AI is getting better, it can search for a broader spectrum of companies and you can ask it for rationale, it will provide you with logic, but are you comfortable fully relying on what the chat window spits back at you?
Can the two be combined?
The appeal of AI search is reach. It can consider a universe a manual screen never would, and catch a genuinely comparable company that simply never surfaced because of the hard filter. It's also a hands-off approach: enjoy your coffee while the tool does the job. The appeal of the filter-driven approach is that every step is visible and repeatable.
Neither has to be sacrificed for the other. A filtered, rules-based universe still needs to exist first, that's what keeps the process auditable, and AI can do real work inside that universe: ranking, flagging near-misses, surfacing candidates a filter alone would have missed. The moment AI is left to construct the universe itself, unconstrained, is the moment nobody can reconstruct how anyone got there.
This isn't just an internal-process question either. Valuation methodology gets real scrutiny in high-stakes settings, fairness opinions, appraisal proceedings, tax disputes, and comparable company selection is very much part of what gets examined when it does.
That's less a settled answer than a genuinely open design question right now. Most tools on the market pick one end of that spectrum. Not many combine them well.
A peer set with a one-line rationale per name (comparable growth profile, matching business model, same geography) survives a partner review. A list of tickers with no reasoning attached doesn't, no matter how it got built.
None of this is a knock on AI-assisted search. It's closer to the opposite. The firms getting real value out of it are the ones treating AI as a research assistant that has to show its homework, not an oracle trusted on vibes.
So back to you. When you build a comps set today, where does the paper trail actually live? In a saved screen, in an analyst's memory, in a chat log with a model, or nowhere at all?
.png){kind=spacer}
